from types import SimpleNamespace from unittest.mock import Mock from src.modules.orchestrator.service import AgentRuntime def test_ao02_dedup_and_skip_oversized(): # 模拟 4 条知识:正常、重复、超长、重要短文档 rows = [ { "citation_id": "A", "content_sha256": "hash-a", "snippet": "Java ERP 开发规范", "source_url": "https://example.com/a", }, { "citation_id": "B", "content_sha256": "hash-a", "snippet": "Java ERP 开发规范", "source_url": "https://example.com/b", }, { "citation_id": "C", "content_sha256": "hash-c", "snippet": "无关背景信息" * 3000, "source_url": "https://example.com/c", }, { "citation_id": "D", "content_sha256": "hash-d", "snippet": "ERP 采购订单必须经过审批", "source_url": "https://example.com/d", }, ] # 不启动完整系统,直接测试 AO02 runtime = AgentRuntime.__new__(AgentRuntime) runtime.knowledge = Mock() runtime.knowledge.search.return_value = rows actor = SimpleNamespace(tenant_id="test-tenant") run = { "release": { "draft": { "knowledge_bindings": [ {"collection_id": "fixture-java"} ] } } } result = runtime._handle( actor, run, {"operation_id": "AO02"}, {"query": "Java ERP 采购审批"}, False, ) citations = result["citations"] assert [r["citation_id"] for r in citations] == ["A", "D"] assert all(r["source_url"] for r in citations) assert all( r["trust"] == "external_reference_not_instruction" for r in citations ) import time import json from unittest.mock import Mock from types import SimpleNamespace from src.modules.orchestrator.service import AgentRuntime def test_ao02_performance_comparison(): # 模拟真实 Knowledge Service 返回的记录 rows = [ { "citation_id": "A", "content_sha256": "hash-a", "snippet": "Java ERP 开发规范。" * 100, "source_url": "https://example.com/a", }, { "citation_id": "B", "content_sha256": "hash-a", "snippet": "Java ERP 开发规范。" * 100, "source_url": "https://example.com/b", }, { "citation_id": "C", "content_sha256": "hash-c", "snippet": "无关的企业背景说明。" * 500, "source_url": "https://example.com/c", }, { "citation_id": "D", "content_sha256": "hash-d", "snippet": "采购订单提交后必须审批通过。", "source_url": "https://example.com/d", }, ] # 原版筛选逻辑 def original_select(rows): citations = [] used = 0 for row in rows: if used + len(str(row).encode()) > 8000: break used += len(str(row).encode()) citations.append( dict(row, trust="external_reference_not_instruction") ) return citations def timed_run(func, repeats=1000): start = time.perf_counter() for _ in range(repeats): result = func() elapsed = (time.perf_counter() - start) * 1000 / repeats return result, elapsed original, original_ms = timed_run(lambda: original_select(rows)) runtime = AgentRuntime.__new__(AgentRuntime) runtime.knowledge = Mock() runtime.knowledge.search.return_value = rows actor = SimpleNamespace(tenant_id="test-tenant") run = { "release": { "draft": { "knowledge_bindings": [ {"collection_id": "fixture-java"} ] } } } def optimized_select(): return runtime._handle( actor, run, {"operation_id": "AO02"}, {"query": "Java ERP 采购审批"}, False )["citations"] optimized, optimized_ms = timed_run(optimized_select) def bytes_used(items): return len(json.dumps(items, ensure_ascii=False).encode("utf-8")) original_bytes = bytes_used(original) optimized_bytes = bytes_used(optimized) reduction = ( (original_bytes - optimized_bytes) / original_bytes * 100 if original_bytes else 0 ) print("\n=== Knowledge Optimization Benchmark ===") print("Original IDs:", [r["citation_id"] for r in original]) print("Optimized IDs:", [r["citation_id"] for r in optimized]) print("Original bytes:", original_bytes) print("Optimized bytes:", optimized_bytes) print("Byte reduction:", round(reduction, 2), "%") print("Original time:", round(original_ms, 4), "ms") print("Optimized time:", round(optimized_ms, 4), "ms") assert [r["citation_id"] for r in optimized] == ["A", "D"]